• My Steam library is 4.3TB.  Less than I was expecting.
  
  Now to move on to backing up our backup server onto my new (old) Synology farm here at PixyLab.  The server contains a historical archive of previous backups that we don't really need online - particularly now that I have fiber internet - and with the crappy COVID-affected exchange rates is costing me more than I really want to spend for something that is a minor convenience.
  
  I also have backups on the new server and on a VPS here in Sydney, and will be moving to encrypted backups on Backblaze B2 now that they are S3 compatible, so that dedicated backup server is really becoming superfluous.
  
  
  
• I've been looking at a company called SSDNodes for a while.  They offer some very good deals on virtual servers - very good deals - if you are willing to pay for 12 or 36 months in advance.
  
  So what's the catch?
  
  Well, they're not a fly-by-night operation, or if they are they're not very good at it; they've been around since 2011.  Multiple reviews say they they oversell CPU (and possibly memory, but definitely CPU) - that is, when you pay for a four core VPS you don't get four cores all to yourself.
  
  Now, that is normal for cheap virtual servers.  If you get an entry-level Lightsail node from Amazon you get one core (because you can't have less) but if you run a sustained load it drops to an effective speed of around 100MHz.
  
  Amazon don't oversell, as such.  Instead they have a system that monitors every single virtual server second-by-second and deliberately chokes the ones using too much CPU time or disk I/O.  But smaller operations like SSDNodes don't have that level of control.
  
  SSDNodes themselves say that their servers aren't overloaded - across all their datacenters they run at an average of 40% load and they balance new VPS creation to prevent hotspots.
  
  Is someone lying here?
  
  No.  In fact, this is exactly what I'd expect.
  
  SSDNodes run on Intel E5 and Gold series Xeon servers, very common and perfectly normal.  Those processors have hyperthreading to improve multi-threaded throughput and turbo boost to improve single-threaded latency.
  
  But the combination of those features completely screws up Linux CPU load figures.
  
  Linux counts each hardware thread as a standalone core for load percentages, but it's smart enough to allocate to individual cores before using the secondary threads.  Since hyperthreading gives - if you're lucky - a 20% performance boost, that means that when you have all cores active - and are seeing a 50% CPU load in your monitoring tool - you are really running at 80% or more of your CPU capacity.
  
  As an added bonus, turbo boost will help increase clock speeds on a lightly loaded CPU, but will spin down as more cores become active.  That means that a server at 40% load is likely really at 80% of capacity already.  And if your true average across all servers is 80%, you will definitely be experiencing hotspots.
  
  So, in short: They are technically not overloaded, but they are pushing the boundary of overselling in a (successful) effort to keep prices down.  You will experience cases where your VPS runs slower than you'd like (particularly because self-similarity likely means that your busy times are the same as other users' busy times).
  
  Fine for some use cases, but definitely not for anything real-time like game servers or video streaming.  For those you need a dedicated server or a mainstream AWS instance with dedicated cores - either of which will cost quite a bit more.
  
  
  
• Gigabyte's B550 Aorus Master has three PCIe 4.0 x 4 M.2 slots.  (Tom's Hardware)
  
  The B550 is an updated X470; it doesn't do PCIe 4.0.  So all three of those must come from the CPU, which means that the main PCIe slot is only 4.0 x8 - at least, if you have more than one M.2 slot active.
  
  Which is still enough to run pretty much any graphics card, so it seems like a reasonable tradeoff for a mid-range motherboard.
  
  Or, alternately, the leak could be wrong and the second and third M.2 slots are PCIe 3.0.
  
  
  
• How well does MongoDB handle distributed transactions?  (Jepsen)
  
  By default, not well.  And if you follow the recommendations rather than the defaults, slowly.
  
  Which is the standard answer for "How well does X handle distributed transactions?" for any X supporting distributed transactions at all.  Unless you have a Tandem Nonstop.  Which I'm guessing you don't.
  
  
  
• The Mac has now been x86-based longer than either PowerPC or 68000.  (Six Colors)
  
  There wasn't much overlap in the PowerPC / x86 switch either.

Posted by: Jay at Saturday, May 16 2020 11:19 PM (0jVI9)

Posted by: Pixy Misa at Sunday, May 17 2020 12:48 AM (PiXy!)

Posted by: Rick C at Sunday, May 17 2020 04:47 AM (Iwkd4)

Posted by: Pixy Misa at Sunday, May 17 2020 02:23 PM (PiXy!)

Posted by: Pixy Misa at Sunday, May 17 2020 07:10 PM (PiXy!)

Hide Comments | Add Comment